GraphGONet

GraphGONet implements a self-explaining neural network that embeds the Gene Ontology (GO) graph into its architecture to predict phenotypes from gene expression data for interpretable biological inference.

Key Features:

• Integration of Domain Knowledge: Embeds the Gene Ontology (GO) graph in hidden layers so that neurons correspond to GO terms representing biological processes, cellular components, and molecular functions.
• Self-Explaining Mechanism: Combines patient-specific gene expression profiles with contextual information from neighboring GO-term neurons to produce interpretable explanations.
• Neuron-level Biological Concepts: Associates individual hidden-layer neurons with distinct GO terms to link network activations to specific biological concepts.
• Interpretability: Produces stable, intelligible explanations that highlight the most influential GO terms and related biological factors contributing to predictions.
• Performance: Reports predictive performance comparable to state-of-the-art non-explainable models while providing explanations.
• Empirical Evaluation: Validated using a microarray dataset (ArrayExpress E-MTAB-3732) and TCGA datasets accessed via the Genomic Data Commons (GDC).

Scientific Applications:

• Phenotype prediction: Predicts disease phenotypes from gene expression data using GO-structured prior knowledge.
• Mechanistic interpretation: Links predictions to GO terms to aid interpretation of underlying biological mechanisms in precision medicine contexts.
• Cross-platform gene expression analysis: Applicable to microarray and TCGA gene expression datasets for comparative or integrative studies.

Methodology:

Constructs a neural network that leverages the structured information of the Gene Ontology by associating each hidden-layer neuron with a specific GO term and integrating patient-specific gene expression profiles with contextual signals from neighboring GO-term neurons to generate predictions accompanied by GO-based explanations.